Context

The company's CEO, Oleg Spirin, approached the Wise IT and gloozd (formerly !FESTCloud.ai) teams with a specific request: to make the search on the platform such that customers could search the way they speak in real life – rather than the way the product catalog is structured.

challenge

Traditional search on VARUS.UA operated under database logic: to find what they needed, customers had to know exactly what to look for. A query like "beetroot" or "chicken fillet" – and the system responded. But if a person wanted to "cook borscht" or "make a light breakfast for kids" – the search failed, leaving the shopper alone with an empty cart and a catalog of several thousand items. This was not just an inconvenient experience. The business was missing out: customers found fewer products than they needed, abandoned their sessions, or left with a smaller receipt. The cognitive load on the shopper undermined conversion – the person had to think for the system instead of the other way around.

Why now?

The online retail market in Ukraine was maturing, and competition was growing fiercer. VARUS.UA aimed not just to "improve search" but to shift the paradigm: to move from SKU-based queries to searching by "life scenarios." A query like "planning a party with friends" had to become just as valid an entry point as "Borges olive oil 500ml." The moment for this leap had arrived – the technology had matured, and the team was ready.

Fears and how we addressed them

Implementing AI into a critical business process – search, through which most transactions flow – always carries risk. Questions like "what if it doesn't work?" and "how much will it cost?" were natural. We suggested starting with a Proof of Concept: first prove the hypothesis in a controlled environment, measure the effect – and only then make a decision about full deployment. This approach removed the pressure of uncertainty. When the PoC confirmed both the technical efficiency and economic viability of the solution (moderate cost with tangible results), the question of "whether to implement or not" disappeared on its own.

solution

The team integrated Vertex AI from Google Cloud into the search logic of VARUS.UA. The heart of the architecture became Gemini – a language model that acts as an intelligent interpreter between the shopper and the catalog. The mechanics are simple and elegant: a customer writes "I want to cook pilaf for four" – Gemini analyzes the query, determines the context, and generates a structured list of relevant products (rice, mutton, onion, carrots, spices). This list is passed to the site's existing search engine, which then returns specific SKUs from the catalog. No infrastructure rewrite – just a smart layer on top. Following a successful PoC, the solution was deployed for all marketplace users.

result

• The numbers speak for themselves.

• Within two months after the launch, the share of AI search among all search sessions grew from 15% to 35% – meaning every third user who opened the search chose the new format completely voluntarily.

• The quality of these purchases proved to be higher: the average value of an item found through AI search is 130 UAH compared to 106 UAH in regular search – a 23% difference.

• AI didn't just help find something; it led to more deliberate, complete purchases.